Akshath Raghav R

I’m a final-year student @ Purdue, majoring in Electrical & Computer Engineering.

I’m passionate about designing solutions that intersect explainable learning algorithms, cloud infra, and HW-SW co‑optimized platforms.

Love to discuss and learn about AI Accelerator Microarch, ML Compilers and Data-Center Infrastructure.

Feel free to reach out to me at araviki[at]purdue[dot]edu.

Fun lil' animation of the Fridrich (CFOP) speedsolving method

background

Smr’25: Interned at BMW (Digital Campus Munich, Germany) in the FG-240 group, working on multiple projects focused on integrating AI Workflows into SDLC Offerings. You can find the projects here.
S’25: Working on a few parameterizable IPs used in the AMP1 Tensor Core, under the AI Hardware Team at SoCET. Completed my Capstone project (through ECE 49022), and won the Senior Design award for the BoilerNet system. Led the backend team under (Beck’s Hybrids)-sponsored project through the Data Mine. Funded by HII to work on a custom FFT-accelerator ASIC.
F’24: Joined the Purdue SoCET Team as a TA. Joined the I-GUIDE Team as a Research Assistant (TDM HDR Fellowship), where we worked on scaling a HPC workflow for distributed inference on Apache Spark (+Sedona) through GCP (accepted at I-GUIDE Forum ‘25).
Smr’24: Joined the Purdue SoCET group, in the Digital Design team. Over the summer, I added the Zicond Extension to the RISCV core for the AFTx07 tape-out. Implemented the TinySpeech family of speech recognition models at VLSI System Design. I also wrote an ANSI-C based inference engine which runs out-of-box in 8bit precision, with 91%+ accuracies for embedded inference.
F’23 - S’24: Worked at the Duality Lab, where we re-engineered the MaskFormer segmentation model (funded by Google!) from the PyTorch-based artifact to TensorFlow for publishing to the TF Model Garden. You can find our paper here and code here. I also generated figures for the PeaTMOSS paper (accepted at MSR’24).
S’24: Led a project at the CVES group @ Purdue ECE, where our goal was to define and evaluate reproducibility within AI/ML projects. I wrote the codebase for building our pipeline and statistically defined the importance of parameters. I was involved in MultiModal (LM) understanding projects at the e-lab. I’ve built eugenie & grammarflow.
S’23 - Smr’23: Interned at Ambee, where I deployed a worldwide fire forecasting system (F3) into their API and wrote automated scripts for their environment-data focused data lakes (still in use). You can find the whitepaper here and access the API here.
F’22 - S’23: Helped lead a project supervised by Prof. Yuan Wang (currently at Stanford) where we aimed to correlate lightning activity with wildfire spread. Wrote (big-)data-interfacing code for satellites across EUR/EUS/SAR, and was responsible for packing them to use within a ConvLSTM model from DeepCube’s short-term forecasting.

Find my reports here.

extras

In my free time, I enjoy photography, chess, reading manga, speed typing (130+ wpm), and whittling!

I can speak English, Hindi, Tamil, and Kannada. Currently, I’m learning German and ASL.

I enjoy Bowling, Golfing, Billiards and Chess. I follow the NBA (Warriors), IPL (RCB) and ATP Masters.

news

Jul 31, 2025	Joined BMW (Digital Campus Munich, Germany) as an intern in the FG-240 group! Working on integrating AI Agents into IDPs and company SDLC.
May 01, 2025	Grateful to receive the “Senior Design” award for BoilerNet.
Aug 06, 2024	Grateful to recieve the Purdue OUR Scholars and DUIRI Scholarships. Excited to be starting as a research assistant in the NSF-funded I-GUIDE team.
Apr 30, 2024	Our team’s report, “A Partial Replication of MaskFormer in TensorFlow on TPUs for the TensorFlow Model Garden,” is now available on arXiv! Find the code here, and report here.
Apr 23, 2024	Received the Outstanding Sophomore in VIP award for the work I did at the CVES group. Read about it here.

latest posts

Aug 06, 2024	[TL;DR] Energy-Based Transferability Estimation
Mar 23, 2024	Set up Llama.cpp on university compute clusters 🦙

papers

A Partial Replication of MaskFormer in TensorFlow on TPUs for the TensorFlow Model Garden

Vishal Purohit , Wenxin Jiang , Akshath R. Ravikiran , and 1 more author

arXiv preprint, arXiv:2404.18801, Apr 2024

Abs

This paper undertakes the task of replicating the MaskFormer model a universal image segmentation model originally developed using the PyTorch framework, within the TensorFlow ecosystem, specifically optimized for execution on Tensor Processing Units (TPUs). Our implementation exploits the modular constructs available within the TensorFlow Model Garden (TFMG), encompassing elements such as the data loader, training orchestrator, and various architectural components, tailored and adapted to meet the specifications of the MaskFormer model. We address key challenges encountered during the replication, non-convergence issues, slow training, adaptation of loss functions, and the integration of TPU-specific functionalities. We verify our reproduced implementation and present qualitative results on the COCO dataset. Although our implementation meets some of the objectives for end-to-end reproducibility, we encountered challenges in replicating the PyTorch version of MaskFormer in TensorFlow.
Time-Driven Fire Risk Forecasting: Leveraging Historical Trends for Enhanced Seasonal Modeling

Roma Jain , Akshath R. Ravikiran , and Pareekshith Katti

May 2023

Abs

Wildfires pose severe threats across North America, causing extensive damage to lives, ecosystems, and property. To address this, accurate fire prediction and forecast outlooks are crucial for effective mitigation. Agencies like the National Interagency Fire Center (NIFC) and the Canadian Wildland Fire Information System (CWFIS) provide vital fire risk assessments. In this paper, our main goal was to demonstrate the sufficiency of historical fire risk data for accurate forecasting. We focused on weather-calculation-based fire risk prediction models, specifically exploring the temporal aspect’s importance in enhancing accuracy. Two encoding methods, One-Shot and Year-By-Year, used for encoding the seasonal changes of fire weather, were analyzed for their implications in fire risk assessment, revealing contrasting attributes. The One-Shot model shows superior accuracy and favorable plots, while the Year-By-Year model offers alternative insights. Despite minor differences in feature importance between the two models, both effectively utilized historical fire weather data for forecasting. This study contributes significantly to fire risk prediction by providing a comprehensive analysis of temporal influences and the effectiveness of different encoding methods. The findings guide model design improvements, bolstering wildfire management and protection measures.